Development of an innovative LSO-SiPM detector module for high-performance Positron Emission Tomography

Transcription

1 Development of an innovative LSO-SiPM detector module for high-performance Positron Emission Tomography Maria Leonor Trigo Franco Frazão Instituto Superior Técnico, Lisboa, Portugal September 1 Abstract Positron emission tomography (PET) is a technique for imaging in nuclear medicine, used to produce functional images, mostly for cancer diagnosis. A PET scanner detects a high energy gamma using scintillation detectors, and the scintillation light using photodetectors. PET scanners for small volumes need a good spatial resolution of the image. One example of this is the ClearPEM mammography scanner, which uses LYSO crystals, with one-to-one coupling to avalanche photodiodes (APDs), and includes depth of interaction (DOI) information from dual-ended readout. The aim of this work was the research on a new detector module with a different configuration: using crystals with smaller section, silicon photomultipliers (SiPM), with a coupling of four crystals to one SiPM (introducing the problem of crystal identification) and, if possible, DOI information from single-ended readout. To accomplish this, different ways of sharing the scintillation light between the photomultipliers were investigated, both with simulations in Geant and experiments. Using a light guide between the crystals and the photomultipliers, the efficiency of crystal identification in simulation was 8.% and in experiment (estimated using a different method for identification) was 95.3%. Using the crystal walls with transparent wrapping for light sharing, the efficiency was in general lower. The dependence of DOI with light sharing was only investigated in simulations. The average DOI resolution obtained for eight selected crystals was 3.8 mm. Keywords: PET, SiPM, ClearPEM, Geant 1. Introduction The present work studied a high resolution detector module for Positron Emission Tomography (PET), in particular for the ClearPEM breast imaging scanner (PET Mammography scanners developed by the Crystal Clear Collaboration) [1, ]. This project was hosted by LIP (Laboratrio de Instrumentao e Fsica Experimental de Partculas) and CERN (European Organisation of Nuclear Research). PET is an imaging technique that uses the distribution of a radioactive marker inside the body of the patients to form an image of the function of the organs in study. The radioactive isotopes used emit a positron, which in turn reacts with an electron in the medium, originating two high energy photons (gamma rays), which are almost anti-parallel and have an energy of 511 kev each. These gammas are detected by scintillating crystals, producing low energy photons that are then detected by photodetectors and converted to an electrical signal which is used to reconstruct the image. The module proposed is composed by mm 3 LYSO:Ce crystals and 3 3 mm silicon photomultipliers (SiPM), with a coupling of four crystals for each SiPM. The SiPM arrays used were Multi-Pixel Photon Counters (MPPC) from Hamamatsu with an active area of 3 3 mm and a pitch of 3. mm. The position of the crystals in a 8 8 matrix relative to the MPPC array is outlined in Figure 1. The purpose of this study was to find solutions to identify the crystal where the gamma interacts, using crystal configurations that allow the sharing of light between several SiPMs. Another target was to find if it is possible to obtain depth of interaction (DOI) information from that light sharing. Examples of studies that accomplished this are [3] (crystal identification from light sharing), [] (DOI information from light sharing) and [5] (both crystal identification and DOI information). Two crystal matrices were tested in Geant simulations and in experiments (Figure ): The Vikuiti Matrix, in which all crystals are separated by Vikuiti foil; The Mylar Matrix, in which some separations 1

2 Figure 1: Layout of the relative position of the crystal matrix (black squares) over the MPPC array (red squares). The gaps between each crystal have.1 mm and between each MPPC. mm (the gaps are not to scale in the picture). Each MPPC is labelled by a letter from A to D and a number from 1 to, which are represented in the figure as a grid. Figure : Schematics of the Vikuiti (left) and Mylar (right) matrices. The crystals are represented by black squares, the Vikuiti foil is in blue and the Mylar foil in yellow. The Vikuiti and the Mylar are glued to the crystals, so there is still a thin glue layer. are with Vikuiti and some with transparent Mylar. Vikuiti TM ESR (Enhanced Specular Reflector) is a thin polymer film from 3M, that reflects 98% of the incident light.. Simulations All simulations were done in Geant, which is a toolkit for the simulation of the passage of particles through matter []. The Geant physics list used was the electromagnetic (GEmStandardPhysics). The processes added were scintillation (GScintillation) and optical processes: absorption (GOpAbsorption), Rayleigh scattering (GOpRayleigh), Mie scattering (GOpMieHG) and boundary process (GOpBoundaryProcess)..1. Vikuiti matrix simulations The output from each MPPC was used to reconstruct the position of the crystal of the interaction, event by event, from the light sharing (either by the glass or the transmission through the Vikuiti). This reconstruction was made using the principle of centre of gravity weighting, or Anger-type calculation [3], which is the weighted sum of the positions of the centre of each MPPC, where the weight is their light output, normalized to the total light output. The coordinates x and y of the position on the plane are calculated for each event as shown in Equations 1 a and b, x = 1 N p i X i, (1a) P y = 1 P i=1 N p i Y i, i=1 (1b) where P is the total light output of the event, N is the number of MPPCs, p i is the light output of MPPC i, and X i and Y i are the x and y coordinated of the centre of MPPC i. After calculating these coordinates for each event, the values are represented in a D flood histogram. The first simulations were made with a glass as a light guide, to share the light from the crystals. The flood histogram obtained from the data from the simulations is shown in Figure 3. Each event is selected from the photopeak of the main MPPC channel. Since the real interaction crystal is known in the simulation, it was then calculated the efficiency of the identification for each crystal as, from the events assigned to a given crystal, the fraction that was correctly identified. This value reflects how correct was the crystal identification. Making the selection only on the photopeak of the main channel, the average efficiency is 8.±.%. This value reflects the effect of the events with Compton scattering, where a considerable energy value was deposited on one of the crystals on the same MPPC channel as the crystal of the first interaction. This can be proven in the simulation by selecting not on the photopeak of the main channel but on the events in which all the energy was deposited on the same crystal. In that case, the average efficiency is 99.±.%. To test if there is a correlation between the light sharing and the DOI, glass plate was placed in the back face of the matrix. It was observed that there is no correlation between the amount of light sharing and the DOI when using polished crystals (as expected from the conclusions of published works, such as [7] and [8]). Therefore, the crystal walls were simulated as unpolished. To keep the simulations as simple as possible, and not depending on many unknown values for those parameters, only the specular spike was used, with a ground finish and a variable σ α. For σ α =.5 and with channel B as an example, the relation between DOI and

3 Y (mm) Flood histogram Entries 195 Mean x Mean y -.19 RMS x.9 RMS y X (mm) 1 Figure 3: Flood histogram of the interaction crystal position using the principle of centre of gravity weighting for the data from the simulations. In this case the light is shared through a 1 mm light guide between the crystals and the MPPCs. light sharing is in Figure (excluding events where there was a Compton scattering from the first crystal to one of the other three over the channel B). DOI (mm) B/total vs DOI Entries 35 Mean x.99 Mean y.1 RMS x.91 RMS y.131 B/total vs DOI B/total Figure : DOI vs light sharing selecting the photopeak of B and excluding the Compton scattering events. A linear fit of this band gives the line y = 9.5x 57.. The histogram of the ratio B/total has a root mean square (RMS) of.91, so the estimated DOI resolution is mm... Mylar matrix simulations The value of transmission provided by the manufacturer of Melinex R 1 is 88.3%. Knowing this transmission value, the index of refraction was calculated to be 1.5. Only a quarter of the Mylar Matrix was analysed, to study the light sharing through the Mylar. First, the MPPC with the highest counts was selected (so the crystal of the interaction was one of the four corresponding to that MPPC), and the desired events are the ones corresponding to its photopeak. Then, two ratios were made in order to select the crystal: R (Figure 5), which compares the light outputs of the two MPPCs adjacent to the main one, in order to select the crystals that are closer to each of the adjacent channels than to the other, and the ones at the same distance; R1 (Figure ), which takes the events from the two crystals at the same distance of the channels adjacent to the trigger, selected in R (the middle peak of R), and compares the light output of the main MPPC to the total light output, in order to distinguish those two crystals. The aspect of the Mylar Matrix that most influences the ratios for the crystal identification is the amount of transmission through the Mylar foil. If there is too much transmission, the ratio R becomes too similar for all the crystals, so the three pretended peaks get overlapped. If the transmission is too low, although the three peaks in R become clearly distinguishable, the two peaks in R1 get closer together. The type of reflection in the Mylar foil is not 3

4 Figure 5: Graphical explanation of the ratio R. The calculation of R is shown by the equation, as the ratio between the output of channel A and the sum of the outputs of A and B1. The R ratio distinguishes between the crystal closer to A than B1, in green (A1.), the one closer to B1, in blue (A1.3) and the two that are at the same distance from A and B1, in red (A1.1 and A1.). This is shown in the three peaks of the plot. The grid on the left shows the MPPC array with the corresponding crystals highlighted in channel A1. Figure 7: Ratio R for transparent Mylar, with index of refraction of 1.5 and diffusive reflection. Figure : Graphical explanation of the ratio R1. The events used for this ratio are the ones in the middle peak of Figure 5. The calculation of R1 is shown by the equation, as the ratio between the output of the main channel A1 and the sum of the outputs of the four channels in the desired quarter of the matrix. The R1 ratio distinguishes the crystals A1.1, which shares less light to the rest of the matrix, in red, and A1., which shares more light, in green. The grid on the left shows the MPPC array with the diagonal crystals in channel A1 highlighted. The mean values of the two peaks can be different than the ones exemplified, depending on the amount of light shared. known, so both diffusive (Lambertian) and specular reflections were simulated. The diffusive surface was observed to allow for too much light transmission and the resulting R is in Figure 7. Figure 8 shows R when using a specular surface, where there is less transmission than in the diffusive. The three peaks are noticeable there. It was concluded from Figures 7 and 8 that less light transmission through the Mylar foil may give a better crystal identification. Figure 8: Ratio R for transparent Mylar, with index of refraction of 1.5 and specular reflection.

5 3. Experiments 3.1. Vikuiti Matrix experiments The experiments with the Vikuiti Matrix were made first without the glass light guide and grease coupling. The purpose is to see if the light sharing between the crystals due to this transmission, and to the optical cross-talk in the grease and epoxy layers, is enough for the crystal identification. The acquisition was made triggering on one of the MPPC channels, and acquiring the other channels at the same time: for each event 1 signals coincident in time were acquired. The events corresponding to the photopeak of the trigger channel are selected. The resulting output spectra are shown in Figure 9, for the trigger channel (C3) and its adjacent channels, where it is possible to observe two peaks clearly distinguishable. The crystal identification can be done either as explained in Section.1 (with an Anger-logic algorithm) or using the MPPCs adjacent to the trigger: the two peaks on the right and left MPPCs distinguish between columns of crystals and the on the up and down MPPCs distinguish between the rows. The former method was not as efficient as the latter for the corner and edge MPPCs, as can be observed in Figure 1. The identification from the adjacent MPPCs is possible even for the corner and edge MPPCs because two peaks are always observed in those adjacent MPPCs. Estimated efficiencies for this method can be seen later. The results of the experiments with a glass light guide are similar to the ones without it. In order to determine which condition gives a better crystal identification, the efficiencies were estimated, using the peaks of the channels adjacent to the trigger. When the trigger MPPC has 3 or adjacent MP- PCs (which happens for all the channels except the corners), it means that in one or the two directions, respectively, there is information from two different channels. These two channels were compared, event by event, checking how many events assigned to one of the peaks in the first channel were assigned to the corresponding peak in the other channel. The value for the division between the two peaks was calculated by middle value = µ 1/σ 1 + µ /σ 1/σ 1 + 1/σ, () where µ i and σ i are the mean and the sigma of the gaussian fit to the peak i. This equation gives the value that separates the same number of events for each side, for two gaussian peaks of the same height. In this case, the peaks are not of the same height, but this should be a good first approximation. The results are summarized in Table 1. The efficiencies with the glass are higher than without it. Table 1: Estimated efficiencies for each trigger channel of the Vikuiti Matrix, with and without the glass. Estimated efficiency (%) Trigger Channel With glass Without glass A 9.±. 85.±. A3 9.9± ±. B1 9.±. 9.1±. B 95.1±. 83.±. B3 93.± ±. B 9.±.3 7.±. C1 97.±. 89.5±. C 9.1±.3 8.±. C3 9.±.3 8.±. C 95.9±.5 8.8±. D 9.8± ±. D3 97.±. 8.±. Average 95.3±.1 8.7±. 3.. Mylar Matrix experiments The first acquisition was done the same way as for the Vikuiti Matrix, but using only channels instead of 1. Figure 11 shows the charge spectra for the simultaneous acquisition of each channel when triggering on A1, and selecting the events corresponding to the photopeak of the trigger channel. In this plot, in channel A, which is adjacent to the trigger, two peaks can be seen, corresponding to each column of crystals over A1. However, in B1, the other adjacent channel to A1, the two peaks are overlapped. This suggests that the crystal identification in the Mylar matrix may not be as efficient as in the Vikuiti matrix. For the Mylar Matrix, some measurements were also made tagging the crystals, in order to select the crystal where the interaction of the gamma occurred. This was done by using as a trigger a PMT that only received the light from one crystal. For that, a black mask was taped to the PMT, with a hole smaller than the crystal section (around 1 mm). This was made for a total 1 crystals and it meant that the back face could not be covered with a Vikuiti foil. The R plots for tagging the crystals over A1 are in Figure 1, as well as their sum. The corresponding R1 plots are in Figure 13, where the sum histogram is the R1 ratio for the events selected in the middle peak of the sum histogram of R. Figures 1 shows the R plot for tagging the crystals over A. It is clear from Figure 1 that this ratio is more overlapped for these crystals than for the crystals over A1. The efficiency of the crystal identification was calculated, for each of the four MPPCs and for each 5

6 Figure 9: Charge spectra the trigger channel C3 and its four adjacent channels of the Vikuiti Matrix cut in the photopeak of channel C Position reconstruction (grease) Entries 5883 Mean x.33 Mean y RMS x.18 RMS y Figure 1: Flood histogram of the Vikuiti Matrix with grease coupling. Because the reconstruction was made for each trigger MPPC separately, the scale of the number of counts is normalized for each of their number of events. group of events assigned to a given crystal. The average efficiencies for each MPPC channel are in Table. As was expected from the examples shown before, the average efficiency for A1 (88.5%) is higher than for A (7.3%). Although the crystal identification is possible with a good efficiency in some cases, in others the efficiency is low. The fact that the ratios calculated are overlapped for some crystals making them difficult to identify suggests that in those cases Table : Average efficiencies of the crystal selection for each trigger channel in the Mylar Matrix. Trigger Channel Average efficiency (%) A A 7.3 B B.5

7 Channel B1 after cut Channel B after cut Counts 5 3 Counts 7 5 Counts ADC Channel Channel A1 after cut - Trigger ADC Channel Counts ADC Channel Channel A after cut ADC Channel Figure 11: Charge spectra of the channels cut on the photopeak of channel A1, with Vikuiti in the back. N 18 1 Sum histogram Crystal A1.1 Crystal A1. Crystal A1.3 Crystal A1. Plot R, trigger on channel A1 r Entries Mean.783 RMS.811 N 5 Sum histogram Crystal A.1 Crystal A. Crystal A.3 Crystal A. Plot R, trigger on channel A r Entries Mean.59 RMS B1/(B1+A) B/(B+A1) Figure 1: R plots for triggering on each of the four crystals over channel A1, and respective sum (in gray). N Sum Histogram Crystal A1.1 Crystal A1. Plot R1, trigger on channel A1 r1 Entries 139 Mean.595 RMS A1/(A1+A+B1+B) Figure 13: R1 plots for triggering on the crystals A1.1 and A1.. The sum histogram is the R1 plot for the events selected in the middle peak of the sum histogram of R, in Figure 1. there is too much light sharing. Figure 1: R for triggering on each of the four crystals over channel A, and respective sum.. Conclusions This work investigated ways of obtaining high resolution in a positron emission tomography module. The module was composed of LYSO crystals and silicon photomultipliers. Simulating the Vikuiti Matrix, it was possible to identify the crystal of the interaction. Using the method of the centre of gravity weighting, the average efficiencies obtained were 85.5% using all the events from the photopeak of the main channel and 99.8% using only the ones in which the gamma ray deposited all the energy in one crystal (and thus excluding the Compton scatterings). Using the same matrix but with the crystal walls slightly unpolished and with a 1 mm glass on top, there is some DOI information. The estimated DOI resolution is 3.8 mm. With the Mylar Matrix, it is only possible to identify the crystals if the light sharing is not too much. It is not possible to obtain DOI information in the 7

8 way that was proposed. From the experiments with the Vikuiti Matrix, it is possible to identify the crystals. Coupling the matrix with the MPPC array only with grease, the average estimated efficiency is 85%. The transparency of the Vikuiti foil, together with the optical grease and epoxy layers, are a source of enough light sharing for crystal identification. Placing a piece of glass 1 mm thick, to work as a light guide, between the matrix and the MPPC array (with grease coupling to both the matrix and MPPCs) gives a better efficiency of crystal identification, with an average of 95%. The crystal identification in the Mylar Matrix has very different efficiency values for each of the MPPC channels. The highest efficiency, 88.% for channel A1, is lower than the identification efficiencies estimated for the Vikuiti Matrix. This might be because the value of the transmission through the Mylar (83%) is too high, making the light sharing from each crystal too similar, like what was seen in the simulations with more light sharing. The clear difference between channels can be due to the way the matrix was assembled, with possible issues such as the size of the Mylar foils cut, or the amount of glue used each time. References [1] Jorge A Neves. The ClearPEM breast imaging scanner. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 8(1): 7, 11. [7] Yongfeng Yang, Yibao Wu, and Simon R Cherry. Investigation of depth of interaction encoding for a pixelated lso array with a single multi-channel PMT. Nuclear Science, IEEE Transactions on, 5(5):59 599, 9. [8] Silin Ren, Yongfeng Yang, and Simon R Cherry. Effects of reflector and crystal surface on the performance of a depth-encoding PET detector with dual-ended readout. Medical physics, 1(7):753, 1. [] G. Cucciati et al. Development of clearpemsonic, a multimodal mammography system for pet and ultrasound. Journal of Instrumentation, 9(3):C38, 1. [3] T. Yamaya et al. A SiPM-based isotropic- 3D PET detector X tal cube with a threedimensional array of 1 mm3 crystals. Physics in medicine and biology, 5(1):793, 11. [] R. S. Miyaoka et al. Design of a depth of interaction (DOI) PET detector module. Nuclear Science, IEEE Transactions on, 5(3):19 173, [5] M. Ito et al. Continuous depth-of-interaction measurement in a single-layer pixelated crystal array using a single-ended readout. Physics in medicine and biology, 58(5):19, 13. [] S. Agostinelli et al. Geant, a simulation toolkit. Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 5(3):5 33, 3. 8